Comparative features for machine learning based classification

Abstract

Systems and methods for generating one or more comparative features for machine learning based classification are disclosed. A system may be configured to obtain time series data and forecast one or more predicted values based on the time series data. The system may also be configured, for each predicted value of the one or more predicted values, to compare an actual value of the time series data to the predicted value and generate a comparative value of a comparative feature based on the comparison. The comparative feature is to be provided to a machine learning model for a classification task associated with the time series data. The classification task may include determining whether one or more data values in the time series data is fraudulent based on the comparative feature.

What is claimed is: 
     
         1 . A computer-implemented method for generating one or more comparative features for machine learning based classification, the method comprising:
 obtaining time series data;   forecasting one or more predicted values based on the time series data; and   for each predicted value of the one or more predicted values:
 comparing an actual value of the time series data to the predicted value, wherein the actual value corresponds to the predicted value; and 
 generating a comparative value of a comparative feature based on the comparison, wherein the comparative feature is to be provided to a machine learning model for a classification task associated with the time series data. 
   
     
     
         2 . The method of  claim 1 , wherein comparing the actual value to the predicted value includes determining a difference between the actual value and the predicted value. 
     
     
         3 . The method of  claim 2 , wherein the comparative value is the difference divided by the actual value. 
     
     
         4 . The method of  claim 1 , wherein forecasting the one or more predicted values is based on one of an autoregressive model or a window function. 
     
     
         5 . The method of  claim 4 , wherein the autoregressive model includes one or more of:
 an autoregressive integrated moving average (ARIMA) model;   a Prophet model; or   an exponential smoothing model.   
     
     
         6 . The method of  claim 4 , wherein the window function for each predicted value is based on one or more of a mean, a minimum, or a maximum of a predefined number of values in the time series data preceding the predicted value. 
     
     
         7 . The method of  claim 4 , wherein the window function includes a naive forecasting model. 
     
     
         8 . The method of  claim 1 , wherein forecasting each predicted value of the one or more predicted values is based on one-step-ahead forecasting. 
     
     
         9 . The method of  claim 1 , further comprising zero filling missing entries in the time series data before forecasting the one or more predicted values. 
     
     
         10 . The method of  claim 1 , wherein:
 the time series data includes daily batch data for one or more merchants ordered by date, wherein the daily batch data includes one or more of:
 a daily count of all monetary transactions for a merchant; 
 a daily count of automated clearing house (ACH) total transactions for the merchant; 
 a daily count of credit card (CC) total transactions for the merchant; 
 a daily count of ACH sales transactions for the merchant; 
 a daily count of CC sales transactions for the merchant; 
 a daily monetary amount of ACH total transactions for the merchant; 
 a daily monetary amount of CC total transactions for the merchant; 
 a daily monetary amount of ACH sales transactions for the merchant; 
 a daily monetary amount of CC sales transactions for the merchant; or 
 a daily batch monetary amount processed for the merchant; and 
   the classification task associated with the time series data includes determining whether one or more daily batches of the daily batch data for the merchant is fraudulent based on the comparative feature.   
     
     
         11 . A system for generating one or more comparative features for machine learning based classification, the system comprising:
 one or more processors; and   a memory storing instructions that, when executed by the one or more processors, causes the system to perform operations comprising:
 obtaining time series data; 
 forecasting one or more predicted values based on the time series data; and 
 for each predicted value of the one or more predicted values:
 comparing an actual value of the time series data to the predicted value, wherein the actual value corresponds to the predicted value; and 
 generating a comparative value of a comparative feature based on the comparison, wherein the comparative feature is to be provided to a machine learning model for a classification task associated with the time series data. 
 
   
     
     
         12 . The system of  claim 11 , wherein the operations for comparing the actual value to the predicted value include determining a difference between the actual value and the predicted value. 
     
     
         13 . The system of  claim 12 , wherein the comparative value is the difference divided by the actual value. 
     
     
         14 . The system of  claim 11 , wherein forecasting the one or more predicted values is based on one of an autoregressive model or a window function. 
     
     
         15 . The system of  claim 14 , wherein the autoregressive model includes one or more of:
 an autoregressive integrated moving average (ARIMA) model;   a Prophet model; or   an exponential smoothing model.   
     
     
         16 . The system of  claim 14 , wherein the window function for each predicted value is based on one or more of a mean, a minimum, or a maximum of a predefined number of values in the time series data preceding the predicted value. 
     
     
         17 . The system of  claim 14 , wherein the window function includes a naive forecasting model. 
     
     
         18 . The system of  claim 11 , wherein forecasting each predicted value of the one or more predicted values is based on one-step-ahead forecasting. 
     
     
         19 . The system of  claim 11 , wherein the operations further comprise zero filling missing entries in the time series data before forecasting the one or more predicted values. 
     
     
         20 . The system of  claim 11 , wherein:
 the time series data includes daily batch data for one or more merchants ordered by date, wherein the daily batch data includes one or more of:
 a daily count of all monetary transactions for a merchant; 
 a daily count of automated clearing house (ACH) total transactions for the merchant; 
 a daily count of credit card (CC) total transactions for the merchant; 
 a daily count of ACH sales transactions for the merchant; 
 a daily count of CC sales transactions for the merchant; 
 a daily monetary amount of ACH total transactions for the merchant; 
 a daily monetary amount of CC total transactions for the merchant; 
 a daily monetary amount of ACH sales transactions for the merchant; 
 a daily monetary amount of CC sales transactions for the merchant; or 
 a daily batch monetary amount processed for the merchant; and 
   the classification task associated with the time series data includes determining whether one or more daily batches of the daily batch data for the merchant is fraudulent based on the comparative feature.